Open-HumanWAM open-sources 50K-hour egocentric robot recipe
TL;DR
- Open-HumanWAM pretrains on 50,000 hours of egocentric human video and reports a log-linear link between pretraining hours and robot success rates.
- HumanWAM-Vanilla reaches 92.08% on RoboTwin 2.0 Clean2Clean and 68.88% on Clean2Random; the goal-conditioned variant lifts Clean2Random to 84.60%.
- The authors release the 50K-hour corpus, data curation and pretraining pipelines, model checkpoints, inference code, and simulation evaluation suites.
The first fully open recipe for scaling egocentric robot pretraining to 50,000 hours, Open-HumanWAM, reports an empirical log-linear scaling law linking pretraining data hours to downstream robot success rates.
On RoboTwin 2.0, the human-only pretrained HumanWAM-Vanilla reaches 92.08% success on Clean2Clean and 68.88% on Clean2Random, alongside a 14.17% mean success rate and a 19.00 mean progress score on RoboDojo. A joint human-robot variant, HumanWAM-Goal, pushes Clean2Random to what the paper calls "a substantial leap to 84.60% on Clean2Random." On physical robots, the authors report the goal-conditioned model "exhibits emergent embodied in-context learning, dynamically adapting its manipulation behavior to unseen visual goals without parameter updates."
The release bundles the Open-HumanWAM-50K corpus, which the abstract describes as "densely annotated with hand motion and natural language," along with the data curation and pretraining pipelines, pretrained and post-trained checkpoints, inference code, and simulation evaluation suites. The headline numbers come from simulation; the abstract gives no quantitative measurement of the physical-robot in-context-learning claim beyond the qualitative description.
Originally reported by paper
Read the original article →Original headline: Open-HumanScale: First Fully Open Egocentric Robot Pretraining Recipe Proves Log-Linear Gains, Scales to 50K Hours